Understanding the risk of diabetic retinopathy from glucagon-like peptide-1 receptor agonists: a Mendelian randomization study and systematic review of European populations

Abstract

Background

Glucagon-like peptide-1 receptor agonists (GLP-1RA) are extensively prescribed to treat obesity and diabetes. However, previous studies have debated whether GLP-1RA use induces diabetic retinopathy (DR) in diabetic populations, and the relationship between the two is unclear.

Methods

Cis-expressed quantitative trait locus data (Cis-eQTL) in blood tissue were used to extract single nucleotide polymorphisms (SNPs) as a genetic proxy tool. The analyses were performed using Mendelian randomisation (MR) as the primary tool and Summary-data-based Mendelian Randomization (SMR) as an auxiliary validation. Discovery cohort was obtained from a large study from the GWAS catalog database, and the FinnGen consortium DR data were used as a validation cohort. Additionally, the outcomes of the two cohorts were combined using meta-analysis. In addition, we systematically retrieved relevant cohort studies of GLP-1RA and DR for systematic review to complement the association of GLP-1RA with DR in the real world.

Results

A total of 9 SNPs highly correlated with the exposure were screened as tool variables to proxy for GLP-1RA. The MR method showed a significant association between GLP-1RA and reduced risk of DR (OR = 0.59, 95%CI: 0.39–0.89, P = 0.0109), in addition, similar results were also found with the SMR method (OR = 0.48, 95%CI: 0.27–0.86, P = 0.0129). Finally, a total of three eligible articles were included in the systematic review, and overall GLP-1RA reduces the incidence of DR compared with existing glucose-lowering agents, but more research is required to verify the generalisability of the findings.

Conclusion

Based on MR and SMR, we found that GLP-1RA can reduce the risk of DR. Systematic review showed that compared with insulin therapy, T2D patients treated with GLP-1RA had a lower incidence of DR, but compared with other hypoglycemic agents, the incidence of DR was inconsistent. Therefore, clinical trials with larger sample sizes and longer follow-up times are warranted to determine this.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13098-025-01878-3.

Introduction

Globally, the number of people with diabetes continues to increase quickly due to a combination of hereditary and environmental factors, with some research claiming that diabetes may influence almost 693 million people globally by the year 2045. However, common microvascular complications such as the threatening diabetic retinopathy (DR) have not only exacerbated the condition of diabetic patients but also significantly reduced their prognosis, and by 2020, the number of people with DR had exceeded 100 million globally, and is still on the rise [1, 2]. Although medications for diabetes mellitus have been updated and evolved in recent years, it is not clear whether the use of these drugs in patients with diabetes induces DR, therefore, it is especially crucial to thoroughly examine the safety of these drugs.

As a leading medication approved by the U.S. Food and Drug Administration (FDA) for the treatment of individuals with type 2 diabetes (T2D) and obesity, glucagon-like peptide-1 receptor agonists (GLP-1RAs) such as semaglutide and liraglutide significantly improved body weight in patients with T2D while reducing glycated haemoglobin (HbA1c) levels in these patients [3, 4]. Moreover, in the Multicentre Long-Term Cardiovascular Outcomes Trial, GLP-1RA decreased all-cause mortality, vascular mortality and non-fatal heart attacks in participants [5]. It is worth noting that despite the potential benefits of GLP-1RA, evidence from the SUSTAIN-6 clinical trial suggests that GLP-1RA may be a risk factor for DR, whereas the clinical trial from LEADER demonstrated that the two are not related [1, 6]. As a result, it is crucial to investigate the precise relationship between GLP-1RA and DR using innovative research techniques.

As genome-wide association studies (GWAS) continue to evolve, the Mendelian randomisation (MR) method provides a novel perspective on the study of causality between two traits from a genetic direction. MR is a method that employs single nucleotide polymorphisms (SNPs) related to clinical phenotypes as tool variables to probe underlying causality between exposures and outcomes of interest [7]. Furthermore, based on MR, Yang et al. [8] created the Summary-data-based Mendelian randomisation (SMR) method and the related instrumental heterogeneity test (HEIDI) method for assessing the robustness of SMR results, a great achievement that enriched the use of MR method further. In conclusion, compared with existing observational studies, MR with SMR simulates the process of random assignment of parental alleles to offspring, and thus can effectively reduce confounding-induced bias and enhance causal inferences [7].

Consequently, in the context of this research, we used cis-expression quantitative trait locus (Cis-eQTL) data of the GLP-1R gene in blood tissues instead of GLP-1RA and assessed the risk association between GLP-1R gene expression and DR using MR as the primary analytical method while referring to the auxiliary analytical results of the SMR method, and to reduce the bias introduced by caste, we limited the participants to be of European origin. We also systematically searched for cohort studies based on European populations to further analyse the association between GLP-1RA and DR in the real world.

Materials and methods

Research design

A drug-targeted Mendelian randomisation approach was utilised to investigate the causal link between increased GLP-1R gene expression and DR in European populations using SNP pooled data of the GLP-1R gene as an instrumental variable. In addition, we used GWAS data of HbA1c, T2D for double positive control to test whether the instrumental variables were representative and reliable. Meanwhile, we used European population data from two different databases as a discovery cohort and a validation cohort to analyse the association of instrumental variables with DR, respectively. In addition, to reduce the bias caused by the data sources and to improve the statistical efficacy, we performed Meta-analysis of the two results with the total effect as the final result. The SMR method was used in our auxiliary analyses to test the accuracy of the results. Finally, we comprehensively searched and reviewed clinical cohort studies based on European populations to complement the real-world risk associations between GLP-1RA and DR. Its research process is shown in Fig. 1. The data used in this paper were obtained from publicly available databases and written with reference to STROBE-MR, so there are no ethical issues or conflicts of interest.

Fig. 1.

Flowchart of the study design. Stage1 is the MR/SMR analysis process, and Stage2 is the PRISMA flowchart for a systematic review

Mendelian randomisation analysis method

Single nucleotide polymorphism instrumental variable screening

The eQTLGen Consortium [9] (www.eqtlgen.org) contains blood gene eQTL data from 31,684 human individuals, providing a new platform for the investigation of the correlation between gene expression and inheritance. Therefore, we used GLP-1RA as an exposure factor, and Cis-eQTL summary data (chromosome: 6; gene region: 39016557-1000 kb-39059079 + 1000 kb) from the eQTLGen Consortium for SNPs in the target gene GLP-1R were selected as instrumental variables for genetic proxies. To obtain more robust instrumental variables to improve statistical efficacy, and to fulfil the requirement that instrumental variables are strongly correlated with exposure, we included SNPs with P < 1 × 10^− 5 for further screening [10]. Taking into account the effect of linkage disequilibrium (LD) on the results, a genetically independent SNP clumping threshold of r² < 0.3 was set based on the 1,000 Genomes European Population dataset with a window of 10,000 kb, and an F-statistic (beta²/se²) > 10 was used as the criterion for strongly correlated instrumental variables. In addition, minor allele frequency (MAF) > 0.01 is considered a common rather than rare variant and was also used as a screening criterion for SNPs [11]. The LDlink website [12] (https://ldlink.nci.nih.gov) provides information on confounding traits such as smoking, alcohol consumption, and obesity that are associated with the traits under study. We used this website to query confounders for SNPs and excluded instrumental variables that clearly had confounding effects on outcomes to reduce bias due to confounders.

Instrumental variable double positive control validation

To validate the validity and reliability of the instrumental variables, we extracted data from two large authoritative databases for double-positive control analyses. The UK Biobank database [13] (www.ukbiobank.ac.uk) collects information on whole genome sequence data, health data, and life data for nearly half a million adult UK residents from 2006 to 2010 and reports the participants’ baseline data characteristics. We obtained HbA1c genetic association data (Code:30750) from this database for 344,182 participants as the first positive control outcome. The DIAGRAM Consortium database (www.diagram-consortium.org) contains a large number of GWAS data on patients with T2D, and we acquired the T2D data as a second positive control endpoint from a large meta-analysis of this database that included 80,154 European patients and 853,816 European normal populations [14]. Causal associations of instrumental variables with the two outcomes were analysed using a two-sample MR approach.

DR data collection

Discovery cohort data were extracted from the GWAS catalog database (www.ebi.ac.uk/gwas/) of a UK population-based GWAS information of 1,339 DR patients and 396,859 controls (Code: GCST90435714) [15]. In addition, to further test the reliability of the results and to reduce the error due to sample overlap, we performed validation using the latest data from the FinnGen (www.finngen.fi) database, version R11. The data of the selected validation cohort contained 3,283 patients and 71,585 controls (Code: DM_RETINOPATHY_STRICT). Both of the above DR datasets meet the definition of ICD-10. The relationship between all the loci of the chromosomes of the two datasets and their corresponding P-values is shown in the Manhattan plot, the red dotted line stands for P = 1e-5 (Fig. 2).

Fig. 2.

Manhattan plot of DR’s GWAS data. The abscissa indicates the chromosome position where the SNP is located, the ordinate indicates the P-value corresponding to the SNP, and the red line represents the threshold P = 1e-5. (A) GWAS data from the Manhattan plot of the GWAS catalog database. (B) Manhattan plot of GWAS data from FinnGen database

MR analysis

Since the random-effects inverse variance weighting (IVW) analysis is more accurate and robust [16], this study used random-effects IVW as the main analysis method to compute the odds ratio (OR) and its 95% confidence intervals, and four methods, namely, MR-Egger regression, Weighted median, Simple mode, Weighted mode as reference. Bayesian Weighted Mendelian Randomization (BWMR) was first proposed by Zhao et al. in 2019. It introduces weighted strategies through a Bayesian framework and systematically integrates pleiotropy modelling to address the pleiotropy issues inherent in traditional MR methods [17]. In this study, the BWMR model is introduced as a supplementary analysis to MR methods. Finally, we pooled the MR results from the two databases by Meta-analysis to enhance the statistical efficacy. The analysis was conducted utilizing the ‘TwoSampleMR’ package and ‘BWMR’ package within R software (version 4.3.3). We used the Bonferroni method for multiple corrections and considered P < 0.025 (0.05/2, 2 representing outcomes from both databases) to be statistically significant, whereas P < 0.05 was regarded as indicative of a possible causative relationship [18].

Furthermore, a series of sensitivity analyses were performed to evaluate the robustness of the findings, including: (1) Cochran Q test using the IVW method and MR-Egger regression was used to analyze the magnitude of heterogeneity (P < 0.05 indicated that there was a large degree of heterogeneity). (2) Horizontal pleiotropy test using the MR-Egger intercept method and MR-PRESSO Global test for selected instrumental variables (P < 0.05 considered that horizontal pleiotropy exists, when other potential pathways due to non-exposure routes may influence the outcome).(3) The MR-PRESSO outlier test was used to determine the presence of outlier SNPs (P < 0.05 indicated the presence of outliers), and the leave-one-out method was used to observe the effect of individual SNPs on the results by excluding SNPs one by one, in order to assess the presence of SNPs that significantly contribute to the bias of the results. The above process was done using the ‘TwoSampleMR’ and ‘MR-PRESSO’ packages.

In addition to sensitivity analyses, we performed co-localisation analyses using the ‘coloc’ package of the R software to assess whether the two phenotypes were mediated by the same genetic variants. In this study, we set the cis region of GLP-1R gene as the co-localisation region (including a total of 235 SNPs), and the tacit a prior probabilities were P1 = 1e-4, P2 = 1e-4 and P12 = 1e-5 (P1, P2 and P12 are the probabilities of SNPs in the co-localisation region with significant chain relationship with the gene expression, the risk of DR, and the two) [19]. Finally, the posterior probability (PP) was obtained from the co-localization analysis and the hypothesis (H0-H4) was judged according to PP. (H0: There was no genetic linkage of the two traits in this region; H1: Only one trait is genetically linked in this region; H2: There is only one additional attribute that is genetically associated with this region; H3: The two traits are related but have different causal variations; H4: Both traits are mediated by identical cause and effect variants)

SMR-assisted analysis

The SMR approach, which uses pooled data from GWAS and eQTL for analyses, can to some extent reflect associations between gene expression levels and traits of interest, so we used the SMR approach as an auxiliary analysis to further complement the MR results. We selected 1000 Genomes European population data as the LD reference information, SNPs with P < 1 × 10^− 5 in the cis region of the GLP-1R gene from the eQTLGen Consortium database as the instrumental variable dataset, and utilised the previously selected DR datasets from the GWAS catalog and FinnGen databases as the outcome. As this study used the SMR method for two hypothesis tests (DR data from both databases), after the Bonferroni method of correcting for multiplicity, it was considered that P < 0.025 (0.05/2) represented evidence of significance. In addition, we performed a HEIDI test to assess whether the association was mediated by pleiotropy and LD (P < 0.05 indicates unreliable results). All the above procedures were done based on SMR software (version 1.3.1).

Systematic review of clinical cohort studies

The article selection criteria and retrieval strategy

To gain a more comprehensive understanding of the association between GLP-1RA and DR, we systematically searched Pubmed (https://pubmed.ncbi.nlm.nih.gov), Embase (www.embase.com) and Cochrane library (www.cochranelibrary.com) databases and set the below included and excluded standards with reference to the PRISMA flowchart. Inclusion criteria: (1) Study type: cohort study. Because cohort studies are more consistent with the causal relationship of the etiological chain and have a high degree of argumentation [20]. (2) Study population: European adult population with diabetes (including type 1 and type 2 diabetes). (3) Intervention: GLP-1RA in the exposed group and placebo or other glucose-lowering agents in the controls. (4) Outcome indicators: whether DR occurred or not. Exclusion criteria: non-English language literature, duplicate publications, literature with inaccessible information, incomplete data, and literature from non-European populations.

According to the above inclusion and exclusion criteria, we formulated a search strategy combining free and Medical Subject Headings (MeSH) terms, both of which were retrieved from database construction to 14 August 2024. Details of the search strategies have been shown in Supplementary Table 1.

Selection of articles and quality assessment and information extraction

2 researchers (Shen and Wang) independently sifted the documents, abstracted the information and cross-checked it. Controversial issues were discussed or consulted with a third party for resolution. Data extracted included: first author, year of publication, type of study design, country, sample size, age, period of study, length of follow-up, number of people exposed and unexposed, outcome indicators of interest, and outcome data. In addition, risk of bias was assessed for the screened documents using the Newcastle-Ottawa Scale (NOS), with outcomes expressed as scores.

Results

Information on instrumental variables of genetic agents

Within the cis region of the GLP-1R gene, we extracted a total of 235 SNPs from the eQTLGen Consortium (Supplementary Table 2), and screened 10 SNPs that were prominently linked to exposure factors with a threshold value of P < 1 × 10^− 5 to serve as genetic proxies for the GLP-1R gene (see Supplementary Table 3 for more information). Nevertheless, upon searching for confounders on the LDlink website, we found that the ‘rs10305420’ tool variable was related to the smoking trait, which is regarded as a DR risk factor [21], and thus we excluded this tool variable to reduce its interference with the results. Table 1 shows the data of 9 SNPs that were finally included and all of them had MAF > 0.01 and F-statistic > 10, which demonstrated that no bias caused by weak tool variables was present in our study.

Table 1.

Information on GLP-1RA instrumental variables for genetic agents

SNP	CHR	EA	OA	P	Beta	Se	MAF	F
rs1018437	6	C	T	2.02E-13	-0.061	0.008	0.437	53.989
rs10305439	6	C	A	6.15E-07	-0.042	0.008	0.427	24.866
rs114977861	6	T	C	2.07E-10	0.694	0.109	0.013	40.399
rs11754911	6	A	C	1.43E-06	0.095	0.020	0.049	23.237
rs148605442	6	G	A	8.01E-06	0.128	0.029	0.025	19.934
rs1929899	6	G	A	1.57E-06	0.054	0.011	0.163	23.062
rs4714209	6	C	T	4.18E-13	0.069	0.009	0.253	52.560
rs75516857	6	G	A	9.89E-06	0.108	0.025	0.035	19.534
rs9283907	6	A	G	1.52E-18	0.101	0.012	0.151	77.232

SNP: single nucleotide polymorphism (instrumental variable); CHR: chromosomal location; EA indicates effector allele; OA indicates other allele; P: P value (Significance threshold); Beta: β value; Se: standard error; MAF: minor allele frequency; F: F statistic (Beta²/Se²)

Positive control results

Positive control results (Supplementary Fig. 1) showed that GLP-1RA significantly reduced HbA1c levels (IVW method: OR = 0.95, 95%CI: 0.92–0.98, P = 0.0013), and both Weighted median method (OR = 0.96, 95%CI: 0.93-1, P = 0.0404) and simple mode method (OR = 0.92, 95%CI: 0.86–0.99, P = 0.0486) showed indicative proof of association. In addition, sensitivity analyses indicated that no significant heterogeneity or horizontal pleiotropy was found, suggesting a relatively reliable result.

When we performed a positive control with T2D using the nine tool variables screened, the MR-PRESSO Outlier test found that ‘rs114977861’ and ‘rs1929899’ were outlier instrumental variables (Supplementary Table 4), so we excluded them. MR analysis was conducted using the remaining SNPs with T2D. IVW method revealed that GLP-1RA significantly decreased the incidence risk of T2D (OR = 0.84, 95%CI: 0.74–0.96, P = 0.0082), and Weighted median method also exhibited a similar result (OR = 0.83, 95%CI: 0.73–0.94, P = 0.0028). The above results further determined the effectiveness and reliability of the selected SNPs.

MR insights into the causal relationship between GLP-1RA and DR

As shown in Fig. 3A, in the MR results of the discovery cohort (GWAS catalog), increased expression of GLP-1R in blood tissues corresponded to a significantly lower risk of developing DR (IVW: OR = 0.47, 95%CI: 0.31–0.70, P = 0.0003), and the heterogeneity test results of the IVW method (P = 0.3196) and the MR-Egger regression method (P = 0.6167) indicated that heterogeneity was not significant and could be ignored. The MR-Egger intercept method (P = 0.0882) and the MR-PRESSO Global test method (P = 0.325) both demonstrated that there was no horizontal pleiotropic effect on the causal association of the genetic proxy GLP-1RA with DR, suggesting that the tool variable was not significantly influencing the outcome through pathways other than exposure. Meanwhile, the MR-PRESSO Outlier test method did not find outlier instrumental variables (no data were generated), indicating that the tool variables were overall robust, and the leave-one-out results also found no notable outlier SNPs (Supplementary Fig. 2A), which together with the above results of the sensitivity analyses, suggests that our MR results are relatively reliable.

Fig. 3.

Main analysis results of the MR method. (A) Forest plot of the six analysis methods. Since the MR-PRESSO Outlier test method did not detect SNPs with potentially large bias in the two cohorts, no quantitative result messages about the MR-PRESSO Outlier test method were generated by the R software, so we did not present them in the figure. (B) Meta-analysis of the outcome of the MR method for the discovery cohort and the validation cohort

Similarly, the results of the validation cohort (FinnGen) likewise supported these findings (IVW: OR = 0.71, 95%CI: 0.52–0.96, P = 0.0246). Moreover, the heterogeneity test and the horizontal pleiotropy test also indicated the absence of significant heterogeneity (IVW: P = 0.9921; MR-Egger: P = 0.9815) and horizontal pleiotropy (MR-Egger: P = 0.9151; MR-PRESSO Global test: P = 0.996). Similarly, the results of the MR-PRESSO Outlier test method and the leave-one-out method did not reveal significant outlier SNPs (Supplementary Fig. 2B). Meanwhile, Fig. 3 demonstrates that in the discovery cohort (GWAS-Catalog), the BWMR analysis revealed a statistically significant inverse association between GLP-1R gene expression and the risk of DR development (OR = 0.39, 95% CI: 0.24–0.62, P = 0.0001). Consistent with these findings, the BWMR analysis in the validation cohort (FinnGen Consortium) showed suggestive evidence of a protective effect of elevated GLP-1R gene expression against DR risk (OR = 0.70, 95% CI: 0.51–0.96, P = 0.029). These findings provide robust supplementary evidence and independent validation for the MR-IVW analysis.

The findings of the meta-analysis illustrated in Fig. 3B indicated that, although the P-value exceeded 0.1 and the I² statistic was greater than 50%, the presence of heterogeneity could not be disregarded [22, 23]. The random effect model was used to quantitatively synthesize the main MR results (IVW method) of the discovery cohort and the validation cohort. The total effect is nevertheless meaningful (OR = 0.59,95% CI: 0.39–0.89, P = 0.0109), indicating that there is a significant causal association between GLP-1RA and DR risk reduction in the European population. In addition, we calculated the PPs of the five hypotheses by Bayesian co-localisation analysis and used PP.H4/(PP.H3 + PP.H4) > 0.7 as evidence that the two traits, exposure and outcome, shared the same genetic variants [24–26]. It is not difficult to find that the co-localisation results of both cohorts suggest that exposure and outcome share the same genetic variants (Table 2), further proving that our conclusions are relatively reliable.

Table 2.

Co-localisation analysis results for the discovery and validation cohorts

Cohort	PP.H0	PP.H1	PP.H2	PP.H3	PP.H4	PP.H4/(PP.H4 + PP.H3)
GWAS-catalog	1.55E-12	5.67E-01	2.81E-13	1.03E-01	3.31E-01	0.7627
FinnGen	2.60E-12	9.54E-01	9.71E-15	3.52E-03	4.21E-02	0.9228

PP: posterior probability; H0-H4: four co-location analysis hypotheses

SMR insights into the causal relationship between GLP-1RA and DR

In the SMR analysis between blood GLP-1R target gene expression and DR risk, we visualised the SNP distribution and P-value relationship between the eQTL dataset and the GWAS dataset for DR, where the labelled dots are the instrumental variable top-SNPs (usually those with the smallest P-value in the exposure dataset), and the red line represents P = 0.025, above which indicates significant statistical significance, while the opposite does not suggest a significant cause-and-effect relation between exposure and outcome (Fig. 4A, B). It was easy to find that the results from the discovery cohort showed evidence of a significant association (OR = 0.24, 95%CI: 0.08–0.74, P = 0.013), whereas the results from the validation cohort, although suggestive of an association between GLP-1R expression and a low risk of developing DR, were not statistically relevant (OR = 0.61, 95%CI: 0.32–1.21, P = 0.161), a result that could be attributed to underlying differences in individual samples. Therefore, in order to reduce the possible confounding bias while observing the total effect of the SMR method in the two datasets, we conducted Meta-analysis of the above two results. Since I² < 50% and P > 0.1 satisfy the conditions for the use of the fixed-effects model [22, 23], we used the fixed-effects model to synthesise the two results. The total effect suggested a remarkable correlation between GLP-1R target gene expression and DR risk reduction (OR = 0.48, 95%CI: 0.27–0.86, P = 0.0129) (Fig. 4C).

Fig. 4.

Main analysis results of the SMR method. (A, B) Visualisation of the P-values of all available SNPs (grey dots) for DR in the GWAS catalog database and FinnGen database with the P-values of all SNPs (red or blue graphs below) in the eQTL data and the P-values of the SMR method (diamonds), where top-SNP points above the threshold line (red dashed line) represent the presence of a significant association. (C) Meta-analysis of the results of the two SMR methods. (D, E) Relative effects of SNPs used for the HEIDI test versus SNPs from the eQTL data. The orange dotted line represents the estimation of the SMR correlation effect size for top-SNPs

In addition, the HEIDI test showed no remarkable heterogeneity or pleiotropy that would cause the results to be unreliable in the SMR analyses of the two DR endpoints (GWAS catalog: P = 0.3827; FinnGen: P = 0.9305). Figure 4D, E illustrates the SNP effect sizes of the SNPs used in the HEIDI test (from GWAS catalog or FinnGen) relative to the SNP effect sizes from the eQTL data, and the closer these SNP distributions are to the orange dashed line represents a more precise HEIDI test result. We provide more information on the SMR analyses in Supplementary Table 5.

Insights from systematic reviews of clinical cohort studies

From the three databases, we retrieved a total of 1,508 relevant publications, including Pubmed (149), Embase (1,281) and Cochrane library (78), and the remaining 1,265 articles were included in the primary screening by removing duplicates of published articles. By reviewing the article titles and abstracts, 115 studies were selected for rescreening, resulting in 3 studies for inclusion in the review. The fundamental characteristics and quality assessment outcomes of the studies are presented in Supplementary Table 6. Given the restricted number of studies included and the lack of uniformity in the interventions across these studies, this research employed qualitative analysis to investigate the relationship between GLP-1RA and the risk of developing DR in real-world settings.

The primary results of the three cohort studies are presented and summarised in Table 3. Overall, a total of two studies reported a clear negative association between GLP-1RA and DR [27, 28]. Among them, Zheng et al. pooled T2D patients from several national registries in Sweden and applied COX regression models to explore the association between GLP-1RA and DR using liraglutide, dulaglutide, and semaglutide as exposures, and metformin and sulfonylurea as controls. After adjusting for confounding factors such as age and educational level at the time of diabetes diagnosis, it was found that the risk of DR incidence in the exposed group was significantly lower than that in the control group (HR = 0.44, 95% CI: 0.3–0.64), and this association remained statistically significant after controlling for diabetes duration, hypertension, cardiovascular events, and the doses of metformin and sulfonylurea drugs (P < 0.0001). In subgroup analyses, similar results were observed in men (HR = 0.34, 95% CI: 0.21–0.55), but no statistically significant evidence was found in women (P = 0.0829). Interestingly, this article also analysed the correlation between pancreatic tissue GLP-1R expression and DR risk using the SMR approach and found that the correlation remained significant in background DR versus severe non-proliferative DR. However, the article did not provide detailed information on HbA1c and blood glucose control, which makes it difficult to investigate the impact of GLP-1 RA on blood glucose control and HbA1c in European DR patients, as well as the influence of blood glucose on the results. Antonios Douros adopted a time-varying exposure definition to align with the dynamic nature of drug therapy for type 2 diabetes, using a COX hazard ratio model to conduct a cohort study of T2D patients from the UK Clinical Practice Centre. Exenatide, liraglutide, and lixisenatide were used as single agents or in combination with non-insulin antidiabetic drugs as exposures, with insulin as the control. The study found that the overall risk of DR was reduced with GLP-1RA (HR = 0.67, 95% CI: 0.51–0.9). In subgroup analyses by duration of drug exposure, a significant reduction in the risk of DR was found after one year of drug use, yet, the risk was similar in the exposed and control groups when the duration of drug use was less than 6 months and 6.1–12 months. The study also reported no significant association between GLP-1RA and overall DR incidence compared to two or more currently used oral hypoglycaemic agents, although HR < 1 was found for both duration of use less than 6 months and use longer than 12 months, but still without statistically significant. Moreover, Jakob Hasselstrøm Jensen et al. [29], based on a Danish population-based study, found that using metformin combined with a dipeptidyl peptidase 4 inhibitor (DPP-4I) as a baseline control, this treatment regimen was observed to be associated with an increased risk of DR in the group of GLP-1RA (Exenatide, Liraglutide, Lixisenatide, Dulaglutide) combined with metformin, but relative to the metformin combined with insulin group, this regimen had a lower risk of DR but still higher than that of the metformin combined with DPP-4I group and the sodium-glucose cotransporter protein-2 inhibitor (SGLT-2I) group. Unlike other large-scale real-world studies such as SUSTAIN-6, LEADER, and REWIND [30], which treat DR as a secondary outcome rather than a primary outcome, these three studies directly set DR occurrence as the outcome while conducting retrospective analyses using large-scale cohort studies, thereby enabling a more precise assessment of the relationship between GLP-1RAs and DR. However, these studies also have certain limitations, such as data primarily sourced from registration centres, which may include data errors. Additionally, the interventions in these studies included various GLP-1RAs and other antidiabetic medications, and the controls were selected from currently first- or second-line antidiabetic medications rather than placebos. Therefore, the possibility of drug interactions interfering with the results cannot be ruled out. Compared with other antidiabetic medications, these studies can only indicate whether the effects are superior or inferior to those medications, and cannot directly reveal whether GLP-1RAs can reduce the risk of DR onset. Furthermore, the duration of diabetes in the patients included in these studies was relatively short. An observational study [31] spanning 28 years and involving 4,513 patients with type 2 diabetes revealed that the incidence of DR significantly increased with the duration of diabetes. The incidence of DR in the early stages (0–5 years) was 6.6%, and by 10–15 years, this proportion increased to 24.0%. Therefore, if the duration of diabetes and the follow-up observation period of the included study subjects are insufficient, it may be difficult to determine whether the absence of DR during the observation period is due to the protective effect of the drug or individual factors, potentially leading to false-positive or false-negative results. Collectively, these results indicate that for individuals with T2D, GLP-1RA may be safer and offer potential benefits for retinal progression in these patients compared to traditional glucose-lowering agents such as insulin; however, additional large cohort studies focusing on European populations are necessary to substantiate this perspective.

Table 3.

Main results of three cohort studies based on European populations

Study	Exposure group		Control group
	Events	Nonevents	Events	Nonevents
Deqiang Zheng 2023 [26]	31	2,359	398	11,331
Antonios Douros 2018* [27]	173	271	2,386	8,045
	98	2,508	226	5,330
Jakob Hasselstrøm Jensen 2024 [28]	84	3,946	106	8,847

*The article contains two cohorts: the first row represents GLP-1RA with two and more glucose-lowering drugs (without insulin); the second row represents GLP-1RA with insulin

Discussion

Diabetes mellitus, as one of the most common chronic endocrine diseases, not only influences the normal life of nearly 508 million people around the world, but also has a significant effect on social and economic development [32]. Although the guidelines and treatments for diabetes management are continuously renewed, the more serious problem is that diabetic patients are also at an increasing risk of developing chronic complications such as nephropathy, neuropathy and retinopathy [33]. Among them, DR, as a major ocular complication in diabetic patients, has a poor prognosis and remains a global health concern. However, thanks to the popularity of optical coherence tomography (OCT) imaging and the refinement of anti-vascular endothelial growth factor (VEGF) therapeutic strategies over the past decades, landmark advances for the prevention and management of DR have been achieved [34]. However, in recent years, with the use of novel glucose-lowering drugs such as GLP-1RA [35], which have been a boon for diabetic patients, it is necessary to examine whether these novel drugs carry a risk of inducing DR. Therefore, from a genetically inherited perspective, we used MR and SMR methods, supported by the huge data from public databases, to dig deeper into the association between GLP-1R gene expression in blood tissues (mimicking the effect of GLP-1RA) and the risk of DR in European populations. At the same time, we also attempted to comprehensively search for real-world cohort studies set in European populations and summarise these studies’ perspectives on the relationship between GLP-1RA and the risk of developing DR. We found that both the discovery cohort and the validation cohort showed a remarkable negative association between GLP-1R gene expression and the risk of DR in the primary analysis using the MR method, and these results were in agreement with the Meta-analysis results. In the analysis of SMR method, the results of the discovery cohort were similar to those of MR, while the results of the validation cohort were not statistically significant but still suggested a potential negative correlation between the two, and the results of the Meta-analysis still indicated that GLP-1R gene expression in blood tissues was significantly associated with a reduced risk of DR. Taken together, the results of these analyses suggest that GLP-1RA decreases the risk of DR. Among the three clinical cohort trial publications we searched, two studies reported that the prognosis of T2D patients with GLP-1RA was remarkably superior to that of currently used hypoglycaemic agents, and the other study reported that GLP-1RA increased the risk of DR in T2D patients, but the existence of other confounding factors on the results of the above studies cannot be excluded. Therefore, these findings need to be validated in multicentre randomised controlled trials and cohort studies with longer follow-up periods and including more participants of European origin.

In summary, our study reveals that GLP-1RAs have been proven to have a protective effect against DR in European populations at the genetic level, but their reliability and universality in clinical practice still need to be further verified. Currently, diabetic patients mainly control their blood glucose levels through lifestyle interventions, diabetes self-management education and support, and dependence on insulin or oral hypoglycaemic agents (such as metformin, pioglitazone, etc.) [36, 37]. Given the cardiovascular protective effects of GLP-1RAs, the 2025 American Diabetes Association (ADA) guidelines recommend them as the first-line medication for blood glucose management in patients with T2D and advanced chronic kidney disease (CKD) [38]. However, as a chronic complication of diabetes, DR often lacks obvious symptoms in its early stages, making timely diagnosis and treatment challenging. Additionally, GLP-1RAs can protect the retina by inhibiting endothelial cell dysfunction (ECD), a mechanism that is particularly important in the early stages of diabetic vascular complications. Therefore, for diabetic patients, especially those with early signs of retinopathy, GLP-1RAs may be considered as part of a comprehensive treatment plan with long-term follow-up. Previous clinical trials have confirmed that GLP-1RAs have certain protective effects on the cardiovascular system, microvasculature, and kidneys [39]. Additionally, some studies have found that GLP-1RAs can reduce intraocular pressure and the risk of glaucoma [40]. These findings are particularly important for diabetic patients at high risk of DR, macrovascular disease, or kidney disease, suggesting that GLP-1RAs may help prevent or mitigate the combined effects of multiple complications, thereby improving patients’ quality of life. Furthermore, for overweight or obese T2DM patients, GLP-1RAs with weight-loss effects are recommended for monotherapy or combination therapy [38].

Personalised treatment is particularly critical when developing treatment plans. Based on the results of this study and in conjunction with the patient’s genetic background, clinicians can assess whether the patient is suitable for adding GLP-1RAs to their existing hypoglycaemic regimen to leverage their potential protective effects on retinopathy. In this study, we selected nine SNPs with high association strength with the GLP-1R gene as strong instrumental variables. For European DR patients carrying these genetic variants, GLP-1RAs may be prioritised, with close monitoring of their efficacy in improving retinopathy. For patients without these genetic variants, alternative treatment combinations may need to be explored; however, these inferences require further validation through clinical trials. Regarding drug dosage, the maximum clinical trial dose of GLP-1RAs (such as liraglutide) reported in T1D is 1.8 mg/day subcutaneous injection [6], which can to some extent mitigate the impact of severe gastrointestinal reactions on tolerability and treatment adherence. Clinically, the recommended starting dose for liraglutide is 0.6 mg/day, which can be increased to 1.2 mg/day after one week based on clinical response. The recommended initial dose for semaglutide is 0.25 mg/day, which is increased to 0.5 mg after four weeks, followed by another four weeks, until the maintenance dose (0.5 mg or 1.0 mg) is reached [39]. Currently, the primary side effects of GLP-1 RAs are widely recognised as gastrointestinal reactions, such as nausea, vomiting, and diarrhoea, but the risk of hypoglycaemia is low. It is important to note that when used in combination with diabetes medications known to cause hypoglycaemia (such as insulin or glinides), hypoglycaemic events may occur. Therefore, when using these medications in combination, it is essential to monitor the patient’s tolerance and blood glucose levels and adjust the dosage appropriately [41]. In summary, when developing a hypoglycaemic treatment plan, it is recommended that patients participate in the decision-making process with their healthcare team, taking into account the hypoglycaemic efficacy, potential side effects, availability, and economic burden of the drugs. Additionally, regardless of the circumstances, the effectiveness, safety, risk of hypoglycaemia, and overall treatment status of the patient should be continuously assessed [42].

Given the complex heterogeneity of obesity and diabetes, traditional single-drug or single-target treatment strategies often fail to achieve optimal efficacy. To overcome this limitation, integrating multiple metabolic pathways into a single treatment regimen may enhance therapeutic efficacy through synergistic effects of multiple mechanisms. However, the multi-drug combination regimen faces numerous challenges in drug development and clinical research, with relatively cumbersome processes. Given this, developing single-molecule multi-receptor agonists that can simultaneously act on multiple receptors with balanced activity has emerged as a more promising alternative. Compared to using single agonists separately, single-molecule multi-receptor agonists can simultaneously activate multiple signalling pathways, not only maximising therapeutic effects but also reducing side effects and optimising pharmacokinetic properties. Currently, GLP-1-based single-molecule multi-receptor agonists have evolved into five main types, including GLP-1/gastric inhibitory polypeptide (GIP) receptor agonists, GLP-1/glucagon (GCG) receptor agonists, GLP-1/GIP/GCG receptor agonists, GLP-1/GCG/FGF21 (human fibroblast growth factor 21) triple agonists, and GLP-1/GIP/IGF-1 (insulin-like growth factor 1)/GCG receptor agonists. The emergence of these novel agonists not only opens new avenues for the treatment of obesity and diabetes but also provides clear directional guidance for future drug development, preclinical, and clinical research [43–45].

In addition, the pathogenesis of DR is complex, involving factors such as retinal microvascular damage induced by hyperglycemia, vascular endothelial dysfunction, inflammatory responses, and oxidative stress. These factors are intricately intertwined, forming a complex pathological cascade. Therefore, actively seeking strategies to improve microvascular damage, protect vascular endothelial cell function, and reduce inflammation and oxidative stress are important treatment approaches for DR patients [46, 47]. Currently, most studies report that GLP−1RAs can reduce the risk of disease progression in DR patients and exert retinal vascular and neuroprotective effects through the aforementioned mechanisms. He’s team treated diabetic mice (db/db) with GLP−1RAs (Loxenatide and Semaglutide, et al.) and found that while increasing retinal vascular density and branching index in mice, the length and extension area of blood vessels were also improved. Additionally, GLP−1RAs maintain mitochondrial integrity and regulate redox balance in retinal endothelial cells, thereby inhibiting high blood glucose-induced mitochondrial DNA leakage and STING pathway activation. They also downregulate VCAM−1, IL−1β, and VEGF expression while upregulating VE-cadherin and ZO−1 expression levels, thereby reducing inflammatory responses and endothelial cell death [48]. Puddu et al. [49] further noted that GLP−1R is widely expressed in retinal endothelial cells, and its activation significantly inhibits the expression and secretion of multiple inflammatory factors such as TNF-α and IL−1α, reducing immune cell infiltration. Chen’s research found that GLP−1R and SGLT2 expression was significantly downregulated in DR patients and negatively correlated with high expression levels of pro-inflammatory cytokines such as tumour necrosis factor (TNF-α) and interferon (IFN-γ). This suggests that GLP−1RAs may restore the expression of GLP−1R to re-establish the regulatory capacity of retinal cells against inflammatory signals, thereby improving vascular abnormalities and inflammatory states in the progression of DR [50]. These results indicate that GLP−1RAs play a central regulatory role in alleviating diabetes-related immune inflammatory responses. Additionally, GLP−1RAs can enhance endothelial nitric oxide synthase (eNOS) activity and NO production by activating the PI3K/Akt and AMPK-eNOS pathways, thereby improving endothelial-dependent vasodilation in diabetic mice and reducing endothelial inflammation and vascular constriction. GLP−1RAs may also exert anti-inflammatory effects by reducing the production of advanced glycation end products (AGEs) and inhibiting the activation of the NF-κB pathway, thereby effectively preventing the onset and progression of early endothelial dysfunction in diabetic patients [51]. Further studies have reported that GLP−1RAs suppress inflammatory responses by activating GLP−1R in central neurons and regulating the ‘GLP−1-brain-immune axis.’ This activation pathway can attenuate the induction of plasma TNF-α by various Toll-like receptor (TLR) agonists, ultimately achieving an anti-inflammatory effect. Notably, this mechanism does not rely on the hypothalamic-pituitary-adrenal axis (HPA axis) or lipid-mediated inflammatory regulation but instead regulates inflammatory factor expression through neuroendocrine pathways mediated by activation of α1-adrenergic receptors and δ and κ-opioid receptors, thereby exerting an immunosuppressive effect [52].

The protective effect of GLP−1RAs in diabetic retinopathy is also closely related to their inhibition of oxidative stress [53]. Zeng’s research found that when the GLP−1 analog Exendin−4 was administered early in the treatment of streptozotocin (STZ)-induced diabetic rats, the mRNA expression levels of Sirt1 and Sirt3 in the diabetic group decreased by 67% and 51%, respectively, compared to the control group. However, in the diabetic group treated with Exendin−4, the protein expression levels of Sirt1 and Sirt3 increased by 3-fold and 2-fold, respectively, compared to the diabetic group, suggesting that Exendin−4 enhances antioxidant capacity by upregulating NAD⁺-dependent Sirt1 and Sirt3 expression, significantly reducing Reactive Oxygen Species (ROS) levels and cell apoptosis in the early stages of diabetic rat retinas, and promoting the recovery of impaired visual function [54]. Further studies have found that hyperglycaemia can impair the antioxidant defence system by inhibiting Nrf2 activity, while GLP−1RA can reverse this change by activating the Nrf2 signalling pathway to enhance antioxidant gene expression, thereby promoting ROS clearance. This has been identified as a new target for the prevention and treatment of diabetic complications in multiple studies [55]. Another study found that downregulation of GLP−1R in retinal pigment epithelial (RPE) cells leads to increased intracellular ROS production, which activates the endoplasmic reticulum (ER) stress signalling pathway. ER stress further induces increased expression of p53 protein and Bax gene, ultimately leading to increased RPE cell apoptosis. Exendin−4 treatment can inhibit high glucose-induced ROS production, ER stress signalling pathway activation, and p53 expression [56]. Additionally, oxidative stress is an important factor in inducing retinal ganglion cell apoptosis and autophagy. Studies have shown that activation of the GLP−1 receptor downregulates the expression of NADPH oxidase NOX3 and mitochondrial superoxide dismutase SOD2, thereby reducing ROS levels, blocking the oxidative stress-induced upregulation of Beclin−1 expression and increased LC3-II/I ratio, and delaying autophagosome formation [57]. More importantly, GLP−1R negatively regulates autophagy activity through the ERK1/2-HDAC6 signalling pathway. In the retinas of diabetic rats, GLP−1 treatment significantly reduced the expression of phosphorylated ERK1/2 and its downstream HDAC6. HDAC6 is a key regulator of non-selective autophagy, and its enhanced activity promotes autophagosome formation by deacetylating substrates such as α-tubulin. By inhibiting this pathway, GLP−1R blocks HDAC6-induced autophagic overactivation, thereby alleviating Retinal ganglion cells (RGCs) damage and apoptosis [57]. Additionally, GLP−1R regulates mitochondrial autophagy by maintaining mitochondrial homeostasis. In an H₂O₂-induced RGC−5 cell damage model, liraglutide pretreatment not only reduced the expression of Beclin−1 and LC3-II/I but also upregulated PGC−1α to promote mitochondrial biogenesis while inhibiting mitochondrial autophagy pathways mediated by BNIP3L and Parkin. This process effectively improved mitochondrial dysfunction by maintaining mitochondrial membrane potential, reducing mitochondrial swelling, and inhibiting excessive ROS production, thereby alleviating RGCs apoptosis and axonal degeneration [58]. However, further investigation is needed to elucidate the specific regulatory relationships between GLP−1R agonists and different autophagy subtypes (e.g., selective mitochondrial autophagy, endoplasmic reticulum autophagy), and to clarify the differential roles of GLP−1R in different retinal cell types using technologies such as spatial transcriptomics at the single-cell level, to achieve precise intervention in DR treatment. In summary, these findings suggest that GLP−1R can reverse high blood sugar-induced retinal damage by alleviating microvascular and endothelial cell damage, improving mitochondrial function and homeostasis, and inhibiting the cascade reactions caused by inflammatory factors and ROS.

Notably, increasing epidemiological and basic research evidence indicates that the pathogenesis of DR is not limited to microvascular lesions; functional impairments and degenerative changes in retinal neurons, particularly ganglion cell damage, are present in the early stages of the disease. This ‘neurovascular unit damage’ is considered a key pathological basis for early-stage DR [59, 60]. Studies have shown that GLP−1R is expressed in the retina of humans and various diabetic animal models, providing a distribution basis for GLP−1RAs to directly act on retinal neurons [61, 62]. Additionally, the neuroprotective effects of GLP−1RAs have been shown to be achieved through multiple mechanisms. On one hand, GLP−1RAs (such as Exendin−4) can effectively inhibit the expression and activity of L-type voltage-gated calcium channels (L-VGCCs) under hyperglycaemic conditions, thereby reducing calcium overload caused by excessive calcium influx, preventing the activation of downstream neuronal apoptosis pathways. By activating the GLP−1R-mediated Gs/cAMP/PKA/ryanodine/Ca²⁺/CaM/Calcineurin/PP1 signalling pathway, L-VGCCs are inactivated, significantly reducing RGCs loss without significantly affecting blood glucose levels, and enhancing discharge capacity under light stimulation, thereby achieving protection of RGCs [63]. On the other hand, GLP−1RAs can also prevent neuronal apoptosis caused by glutamate excitotoxicity by regulating glutamate homeostasis in the retina. In the db/db mouse model, local administration of GLP−1RAs reverses diabetes-induced downregulation of GLAST (glutamate-aspartate transporter), reduces extracellular glutamate accumulation, and activates anti-apoptotic signalling pathways such as Bcl−2 and AKT, thereby inhibiting the expression of apoptotic factors such as iNOS, FasL, and caspase−8 [62, 64]. Additionally, some researchers have found that local administration of liraglutide via eye drops can effectively replenish reduced GLP−1 levels in the retina under diabetic conditions, activate the GLP−1R/Akt/GSK3β signalling pathway, reduce abnormal phosphorylation of tau protein at Ser396 and Thr231 sites, and alleviate synapse damage and mitochondrial dysfunction induced by these abnormalities, thereby exerting neuroprotective effects [65]. In addition to the above studies, another novel research focus is the improvement of functional impairments in early DR by GLP−1RAs through promoting the release of the inhibitory neurotransmitter GABA (gamma-aminobutyric acid). Studies have found that early-stage diabetes (4 weeks of hyperglycaemia) leads to a reduction in the frequency of GABAergic miniature inhibitory postsynaptic currents in RGCs, while GLP−1 eye drops can significantly increase this frequency, enhance RGCs survival rates, and improve visual function. This process is associated with the activation of the phospholipid-phospholipase C (PI-PLC)/inositol 1,4,5-trisphosphate receptor (IP3R)/Ca²⁺/protein kinase C (PKC) signalling pathway, which promotes GABA release from amacrine cells (ACs) to RGCs and inhibits excitotoxic processes associated with DR [66]. These findings suggest that GLP−1RAs may possess direct neuroprotective functions in the retina independent of their hypoglycaemic effects. Therefore, re-examining the pathogenesis of DR from the perspective of ‘early neuropathy’ provides new targets and theoretical basis for early intervention strategies in DR, and also highlights the potential and necessity of locally delivering GLP−1RAs for ophthalmic therapy.

In recent years, the protective effects of GLP−1RAs on the blood-retinal barrier (BRB) have also garnered attention. Upon activation of the GLP−1R, inhibition of the RhoA/ROCK signalling pathway reduces the phosphorylation of myosin light chain (MLC), thereby maintaining the integrity of tight junction proteins such as Occludin and ZO−1, and protecting the BRB from hyperglycaemia-induced damage [52]. Additionally, GLP−1RAs significantly reduce the expression of MMP−9, VEGF, and ICAM−1 and inhibit the ERK1/2 and AKT/PKB signalling pathways, thereby alleviating inflammatory responses and vascular leakage, further protecting the BRB [52, 67]. These mechanisms collectively not only maintain the integrity of the BRB but also improve retinal function in DR, highlighting the potential advantages of the BRB as a therapeutic target for DR patients.

In real-world research, the findings from cohort studies conducted on European populations have been systematically reviewed in the relevant sections of this article. However, the global evidence regarding the association between GLP-1RAs and DR remains inconclusive. A study involving 47 patients reported that GLP-1RA treatment was associated with a transient worsening of DR in the short term, but longer-term follow-up demonstrated a partial improvement in DR risk with continued therapy [68]. In contrast, a retrospective cohort study of 1,626 individuals of Chinese descent revealed a significantly lower incidence of DR in the GLP-1RA group compared to those treated with insulin or oral hypoglycemic agents. This association persisted in both univariate and multivariate logistic regression analyses, even after adjusting for age and sex [48]. Conversely, other studies have suggested a potential increase in DR risk with GLP-1RA use. For instance, a retrospective cohort study utilizing data from the TriNetX global health research network, conducted by A. Eleftheriadou et al., compared DR risk in patients with type 2 diabetes receiving insulin therapy in combination with either SGLT-2Is or GLP-1RAs. The results indicated that patients treated with GLP-1RAs in combination with insulin exhibited a higher likelihood of developing DR compared to those receiving SGLT-2Is with insulin, a risk profile similar to that observed in the insulin-only control group [69]. However, this study was not a randomised controlled trial but an observational cohort study using electronic health record data from populations in the United States, Europe, and the Asia-Pacific region. While this study design is more conducive to temporal causal inference, it may introduce confounding factors such as regional and lifestyle differences, as well as selection bias, which could affect the reliability of the results. Additionally, both the SGLT-2I and GLP-1RA intervention groups and the control group used insulin, but the results of using GLP-1RA alone were not mentioned in the article. Therefore, it is unclear whether there is an interaction between insulin and GLP-1RA that affects the results. Similarly, a previous large-scale clinical trial (SUSTAIN-6) also indicated that the use of GLP-1RAs was associated with an increased risk of DR compared to the placebo group [39], but this result was based on 3,297 T2D patients from 20 countries, and 83.0% of the patients included in the trial already had cardiovascular disease or chronic kidney disease, indicating that the patient population had a high risk of complications. This high-risk population may respond differently to GLP-1RA than the general population, leading to an increased risk of DR. Furthermore, in the subgroup analysis, the authors only discussed the primary outcomes (cardiovascular death, non-fatal myocardial infarction, or non-fatal stroke) and did not perform a subgroup analysis of the association between semaglutide, placebo, and DR. Therefore, the interference of regional and racial differences on the inference of risk associations cannot be ignored. Furthermore, differing definitions of DR may also account for variations in study results. This study defined vitreous haemorrhage, blindness, or the need for intravitreal medication or photocoagulation therapy as secondary endpoints. Since the primary objective of the trial was to assess cardiovascular safety, this may have led to an imprecise estimation of DR risk. Another large clinical trial by the same team (LEADER) reported similar results, with a higher incidence of retinal lesions in participants receiving subcutaneous liraglutide compared to the placebo group, though this difference was not statistically significant. This outcome may be attributed to the fact that most patients were using antihypertensive drugs, statins, and other medications, which could influence cardiovascular and microvascular complications and thereby affect the efficacy of GLP-1RA on DR. Similarly, this study also treated DR as a secondary endpoint, which may have limited the assessment of the association between GLP-1RA and DR. Another key point is that the median observation periods for these two trials (LEADER and SUSTAIN-6) were 3.8 years and 2.1 years, respectively, which are relatively short. For assessing the risk of chronic complications like DR, longer follow-up periods may be needed to observe the long-term effects of GLP-1RAs. Additionally, some studies have reported that an increased risk of DR is associated with a significant reduction in HbA1c levels within a short period of GLP-1RA use [70], suggesting that shorter follow-up periods may not adequately capture potential changes in DR risk associated with GLP-1RA. Therefore, larger-scale, longer-term follow-up, and more precise clinical trials are urgently needed to comprehensively examine the association between GLP-1RA and DR risk, particularly focusing on the use of first-line antidiabetic drugs (such as insulin) or the combination of multiple GLP-1RAs.

Limitations

Although this study strictly followed the analytical procedures, it inevitably has certain limitations. Firstly, regarding data sources, since all data used in this paper are from publicly available databases, the data quality may be affected by the sample size of the original studies and the limitations of different statistical methods. Moreover, we chose blood tissue eQTLs rather than retinal tissue. Compared to blood tissue, retinal tissue has unique cell types and physiological functions, and its gene expression regulation may significantly differ from that of blood tissue [71]. Therefore, the instrumental variables selected based on blood eQTL data may not accurately reflect the true gene expression and functional roles in the retina. Additionally, gene expression in blood may be influenced by various confounding factors, such as inflammatory factors and metabolites, which may not exist or have different mechanisms in the retina, leading to potential biases in the results [72]. Secondly, our analysis was conducted based on a European population, so the generalizability of our findings to other ethnic groups remains debatable. Furthermore, the drug-targeted MR or SMR results reflect the long-term impact of blood tissue genetic proxies for GLP-1RA on DR, which may not align with the results of clinical trials conducted over relatively limited time periods. Lastly, our research focus was limited to the impact of GLP-1RA on the onset of DR, but we did not address whether the use of GLP-1RA in DR patients would exacerbate or alleviate disease progression. In terms of systematic literature review, only three studies were included. On the positive side, we selected these three studies because they focused on the European population and met other stringent criteria, indicating that they have high representativeness and reference value, which to some extent reduces bias and enhances the persuasiveness of our study. Additionally, through in-depth analysis of these studies, we gained a more detailed understanding of the use of GLP-1RA in the European diabetic population, providing direction and insights for subsequent broader research. However, on the negative side, the limited number of included studies also has certain limitations. These studies, all conducted in European populations, make it difficult to generalize the findings to broader populations and different regions. Moreover, potential differences in sample characteristics may lead to high heterogeneity in the results, making it challenging to synthesize them into a universally applicable conclusion. Future research should include more real-world studies to confirm our findings and pay greater attention to the potential role of GLP-1RA in treating DR patients and its impact on prognosis.

Conclusion

In conclusion, the present study revealed that GLP-1R gene expression in blood tissues (used to mimic the effects of GLP-1R agonists) is associated with a reduced risk of developing DR in a European population using the MR method as the primary analysis tool, and this conclusion is in agreement with the results of the SMR method-assisted analyses. We also conducted a systematic review of cohort studies in the context of European diabetic populations in order to explore the association of GLP-1R agonists with DR in the real world in a more in-depth and comprehensive manner; however, large clinical trials with more participants are needed in the future to explore the association of GLP-1R agonists with the risk of DR and to broaden the target of the study to include individuals who already have DR.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

B.S proposed research ideas and collected data; B.S and W.W wrote the manuscript. Y.G, Z.C, J.H, C.L, Y.L strictly examined and improved the manuscript.

Funding

No Funding.

Data availability

All data generated or analysed during this study are included in this published article [and its supplementary information files].

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.